{"id":"5613f250-a4e3-42f9-ade5-4e7da1424bea","arxiv_id":"2603.27493","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A two-stage pure RL method with an information-gap global view and hierarchical grounding loss makes MLLMs truly rely on precise crops and sets SOTA on high-res VQA under tight token budgets.","lead":"Agent-style multimodal models often crop images but still answer from the full scene. This paper trains them with an information gap and a small grounding reward so they actually use the crop, reaching strong high-resolution VQA scores with far fewer visual tokens.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged selection heuristic.","rationale":"The reader already identified the precise soft spot (Method §4.1 Offline Resolution Selection and the Pixelreasoner/CoF/ThinkLite-VL + 256 VisualProbe mix). The paper’s own preliminary analysis and ablations make the crop-dependence claim falsifiable and partially verified; the remaining risks are empirical re-validation risks, not hidden logical gaps. Therefore no verdict adjustment is warranted: keep CONDITIONAL with the same confidence and the same caveat that others must re-validate the information-gap selection rule.","tokens_in":18768,"tokens_out":476,"duration_ms":5855,"concrete_test":"Hold the Stage-I training mix fixed and re-train three variants that replace the answer-divergence rule with (i) fixed hard downsampling to 1,024 tokens, (ii) random scale selection, and (iii) the original high-resolution global image; then re-measure the Acc delta under GT-crop vs. random-noise replacement on HR-Bench 8K (Lite). If the deltas collapse to the baseline levels of Table 5, the selection heuristic is load-bearing and the transfer claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript under review is the MLLM agent paper (information-gap Stage-I + hierarchical grounding Stage-II), not the SNN UAV tracker named in the arXiv ID. Its central claim—that offline resolution selection forces genuine crop dependence and thereby yields SOTA under both 1,024- and 16,384-token budgets—is already supported by the controlled GT/noise crop replacements (Tables 1 and 5) and the two-stage ablations (Tables 3–4). The reader’s weakest assumption correctly isolates the only load-bearing premise: that the offline “answer-divergence” downsampling rule induces a general test-time habit rather than an artifact of the particular training mix or the selection heuristic itself. No stronger internal inconsistency, missing control, or unstated assumption that would overturn the claim was found. The remaining limitations (256-box Stage-II set, hand-tuned thresholds, single-seed runs) are real but secondary and already reflected in the CONDITIONAL verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper identifies that existing agentic MLLMs (SFT+RL or pure RL) for high-resolution VQA often invoke cropping tools perfunctorily, relying primarily on the global image rather than crop content (validated via GT-crop vs. random-noise replacements in Table 1 under both 16 384- and 1 024-token budgets). It proposes a two-stage pure-RL framework without trajectory supervision: Stage I creates an Information Gap by offline iterative downsampling of the global image until the model’s answer diverges from its high-resolution answer, while still extracting crops from the original high-resolution source, so that accuracy reward forces crop utilization; Stage II adds hierarchical nested bbox annotations on 256 VisualProbe samples plus a composite grounding reward (conditional IoU only when coverage > 0.9, plus L1 corner distance). Trained with GRPO on Qwen2.5-VL-7B, the method reports SOTA accuracy on HR-Bench 8K/4K and V* under both token budgets, improved crop-dependence deltas (Table 5), higher IoU, and lower inference time when using only 1 024 global tokens.","tokens_in":19033,"tokens_out":1191,"duration_ms":23622,"significance":"If the central claim holds—that the offline answer-divergence rule induces a transferable test-time habit of attending to high-resolution crop content rather than an artifact of the selection heuristic or training mix—the work supplies a practical, teacher-free pure-RL recipe that simultaneously raises accuracy and reduces visual-token cost for fine-grained agentic VQA. Strengths include clean controlled replacements (Tables 1, 5, 7), systematic ablations isolating Stage I, Stage II, data source, L1 term and IoU threshold (Tables 3–4, 8–10), public benchmarks, released code, and explicit efficiency comparison (Table 6). These elements make the contribution falsifiable and reusable for energy- or latency-constrained multimodal agents.","major_comments":[{"comment":"§4.1 Offline Resolution Selection: the stopping criterion is defined by answer divergence of the very model being trained. This creates a risk that the selected resolutions are tuned to that model’s particular failure modes rather than inducing a general crop-attention habit. A load-bearing control is missing: re-run Stage I with resolutions chosen by a frozen independent model (or by fixed hard/random schedules already partially ablated in Table 9) and measure whether the GT/noise deltas of Table 5 and the SOTA numbers of Table 2 still appear at test time.","section":"§4.1 Offline Resolution Selection"},{"comment":"§5.1 and Table 4: Stage-II gains rest on only 256 manually annotated hierarchical boxes from VisualProbe. The paper shows that reusing Stage-I data yields no gain, yet provides no scaling curve or cross-validation on annotation quantity/quality. Because the hierarchical IoU+L1 reward is the sole precision signal, it remains unclear whether the reported IoU lifts (Table 3) and final SOTA would survive a larger or differently sourced labeled set; this directly underpins the “precise cropping” half of the central claim.","section":"§5.1 Training Datasets; Table 4"},{"comment":"Tables 1b and 5: even after Stage I the accuracy delta under GT versus noise replacement remains modest (≈10–18 points). The paper interprets this as successful crop dependence, yet the residual performance under pure noise (and the continued utility of the global image under the 16 384-token regime) suggests that the Information Gap only partially overrides the global-image shortcut. A quantitative analysis of how much of the final answer still depends on global tokens after training would strengthen or qualify the “genuine reliance” claim.","section":"Tables 1b, 5"}],"minor_comments":[{"comment":"Throughout: the V* benchmark is rendered as “� �” or “V Star”; replace with consistent “V*” and cite the original paper properly.","section":"Tables 1–6, §5"},{"comment":"Eq. (1) and surrounding text contain garbled symbols (�� �� �����); clean the action formulation so the crop coordinate and image variables are unambiguous.","section":"§4 Eq. (1)"},{"comment":"Figure 3 caption and §4.2: the hierarchical bbox notation (B1 \\subset \to Bn) is never formally defined; a short recursive definition would help.","section":"Figure 3; §4.2"},{"comment":"Table 2 footnotes mix evaluation protocols (temperature averaging, option shuffling); state a single consistent protocol for all compared methods.","section":"Table 2"},{"comment":"§6 and Table 6: report wall-clock time on identical hardware and batch size; the current numbers mix different token budgets without error bars.","section":"§6 Efficiency Analysis"}],"recommendation":"major_revision","confidential_remarks":"The supplied arXiv identifier and abstract describe an unrelated SNN UAV tracker; the actual manuscript is the MLLM cropping agent. Confirm the correct identifier before production. The work is solid empirical ML but sits at the boundary of “agentic MLLM tooling”; ensure the journal’s scope includes pure-RL multimodal agents."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The ID/title mismatch is real: the abstract describes STATrack (fully spiking RGB UAV tracking with AMIM), but the full text is “Learning to Focus and Precise Cropping,” a pure-RL agentic MLLM paper. Everything below is about the manuscript we actually have.\n\nWhat is new is the information-gap construction. Prior pure-RL croppers (DeepEyes) and SFT+RL hybrids already exist; the distinctive move is offline resolution selection that downsamples the global image until the model’s own answer diverges, while still cropping from the original high-res source. That forces the crop to carry information the global view no longer has. Stage-II then adds a small hierarchical bbox set (256 VisualProbe boxes) with a conditional IoU + L1 grounding reward. The preliminary GT-vs-noise crop swaps (Tables 1a/b) cleanly show that earlier agents barely depend on crop content; the Stage-I/II ablations and the 1k-token vs 16k-token tables show the fix works and is efficient.\n\nSoft spots are real but secondary. Stage-II rests on only 256 hand-labeled boxes; reward weights, the 0.9 overlap threshold, and the answer-divergence stopping rule are free parameters; results are single-seed Qwen2.5-VL-7B runs with a few protocol footnotes on baselines. The load-bearing premise is that the offline selection rule induces a general test-time habit rather than an artifact of the training mix. Nothing in the paper overturns that premise, but others will need to re-validate it.\n\nMath and data look ordinary for empirical RL work; citations cover the right agentic and high-res VQA lines. Code URL is given. This is for people building tool-using MLLMs who care about token budgets and genuine tool dependence. It deserves a serious referee. I would bring it to reading group and would cite the information-gap idea if I were working on agentic perception.","headline":"The arXiv ID points to an SNN UAV tracker, but the full manuscript is the MLLM cropping paper; that paper’s real contribution is a clean diagnosis of crop under-use plus a workable two-stage pure-RL fix that delivers SOTA under tight token budgets.","tokens_in":19684,"tokens_out":517,"would_cite":true,"duration_ms":6447,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A two-stage pure RL method forces MLLMs to actually use high-resolution crops instead of ignoring them.","keywords":["multimodal large language models","reinforcement learning","visual cropping","information gap","high-resolution VQA","grounding reward","agentic workflows"],"falsifier":"Retrain the identical two-stage pipeline but always use the original high-resolution global image (no information gap) and measure whether crop-dependence (accuracy drop when crops are replaced by noise, accuracy gain when replaced by ground-truth boxes) and the reported SOTA gains disappear.","tokens_in":19573,"feed_emoji":"🔍","tokens_out":804,"duration_ms":14837,"temperature":0.7,"pith_summary":"Existing agentic multimodal large language models that crop image regions often answer from the full image and treat the crop as a formality. This paper shows that creating an explicit information gap—feeding a carefully downsampled global view while still cropping from the original high-resolution image—trains the model to rely on the crop content. A second stage then uses a small set of hierarchical bounding-box annotations plus an IoU-plus-distance grounding reward to make those crops more precise and compact. The resulting models reach state-of-the-art accuracy on high-resolution visual-question-answering benchmarks under both tight and generous visual-token budgets, and remain faster because they process fewer tokens overall.","feed_headline":"RL gap forces MLLMs to use the crops they make","feed_subtitle":"Downsample the global view until the answer flips; accuracy jumps and token cost falls","key_machinery":"The Information Gap mechanism (offline resolution selection that stops at the first downsampling level where the model’s answer changes) together with the later hierarchical grounding reward that combines conditional IoU and L1 corner distance.","core_discovery":"When the global image is deliberately degraded just enough that the model’s own answers diverge from its high-resolution answers, and crops are still taken from the original full-resolution source, pure reinforcement learning produces agents that genuinely attend to and depend on the cropped regions rather than treating the crop tool as optional confirmation.","pith_inferences":["The same resolution-selection trick could be used as a diagnostic probe for any tool-using MLLM: if replacing the tool output with noise barely hurts accuracy, the model is not really using the tool.","Hierarchical boxes may transfer to other open-ended visual tools (segmentation, depth) where a single tight box is too brittle.","Because Stage-I already produces usable crops, synthetic box data generated by pasting annotated low-resolution objects onto high-resolution backgrounds could remove the remaining manual annotation bottleneck."],"forward_implications":["Agentic MLLMs can be trained without expensive teacher trajectories and still outperform hybrid SFT+RL baselines on fine-grained high-resolution VQA.","Under a 1 024-token global budget the method already beats prior agents that are allowed 16 384 tokens, showing that crop reliance yields better token efficiency.","A few hundred hierarchical box annotations are enough to shrink crop area and token count while further lifting accuracy.","The same information-gap idea can be applied at test time by simply downsampling the global view, giving an immediate accuracy boost even for already-trained agents."],"fun_headline_variants":["Fully spiking nets track UAVs from RGB with target-aware features","STATrack: first full SNN RGB tracker for energy-efficient UAVs","AMIM preserves target semantics in spiking UAV visual trackers","Fully spiking STATrack cuts energy use in RGB UAV tracking","Target-aware SNNs enable low-power UAV tracking without event cameras"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The offline rule that keeps downsampling the global image until the model’s own answer flips will reliably teach a lasting habit of using crop content, rather than an artifact of the particular training images or the selection heuristic itself.","fun_headline_variants_meta":{"raw":{"variants":["Fully spiking nets track UAVs from RGB with target-aware features","STATrack: first full SNN RGB tracker for energy-efficient UAVs","AMIM preserves target semantics in spiking UAV visual trackers","Fully spiking STATrack cuts energy use in RGB UAV tracking","Target-aware SNNs enable low-power UAV tracking without event cameras"]},"model":"grok-4.5","effort":"low","cost_usd":0.004952,"raw_usage":{"total_tokens":1388,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":95,"cost_in_usd_ticks":49520000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":545,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":95,"duration_ms":5728,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T16:54:11.249468+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the identical two-stage pipeline but always use the original high-resolution global image (no information gap) and measure whether crop-dependence (accuracy drop when crops are replaced by noise, accuracy gain when replaced by ground-truth boxes) and the reported SOTA gains disappear.","supporting_citations":[],"review_version":1}